2017 - Alex Shepherd - Deep Learning for Conservation

iNaturalist recently integrated computer vision into their apps, providing automatic species suggestions based on visual and spatio-temporal data (i.e. where things have been found before). This is a presentation on how the app works, how it was built (TensorFlow + inception + our dataset), and challenges encountered when integrating deep learning into the expert community of scientists and trained naturalists.

Video: https://youtu.be/MD8EwRGoS3U

Transcript

1. Deep Learning for Conservation Alex Shepard alex@inaturalist.org California Academy of Sciences
3. What is iNaturalist?
4. 188 0 1900 192 0 194 0 196 0 198 0 2000 2020 2,000 1,000 0 iNaturalist Other iNaturalist Other 93% of all Monarch Butterfly records this decade are from iNaturalist GBIF Monarch Butterfly Records / month iNaturalist generates a significant portion of the world’s biodiversity data
6. iNaturalist is doubling every year
7. 5 million images of wildlife 13k species with > 20 observations Train Val Test iNaturalist Deep Learning Dataset
8. Inception v3 Original Random Crop Flipped Cropped Color distorted iNaturalist Deep Learning Training Pipeline
9. Inception v3 iNaturalist Deep Learning: Prediction
10. http://appstore.com/inaturalist iNaturalist Deep Learning Deployed in iOS
11. Top 1 Accuracy 57% Top 2 Accuracy 66% Top 10 Accuracy 78% Common Ancestor Accuracy 93% Vision Accuracy Curve Full System Accuracy iNaturalist Deep Learning: Results
12. iNaturalist Deep Learning: Future Challenges
13. Thanks! https://inaturalist.org alex@inaturalist.org https://github.com/inaturalist https://github.com/visipedia